paarth-sharma · ~/portfolio
main● ready
── 00 / HERO ──────────────────────────
██████╗  █████╗  █████╗ ██████╗ ████████╗██╗  ██╗
██╔══██╗██╔══██╗██╔══██╗██╔══██╗╚══██╔══╝██║  ██║
██████╔╝███████║███████║██████╔╝   ██║   ███████║
██╔═══╝ ██╔══██║██╔══██║██╔══██╗   ██║   ██╔══██║
██║     ██║  ██║██║  ██║██║  ██║   ██║   ██║  ██║
╚═╝     ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝   ╚═╝   ╚═╝  ╚═╝

███████╗██║  ██╗ █████╗ ██████╗ ███╗   ███╗ █████╗
██╔════╝██║  ██║██╔══██╗██╔══██╗████╗ ████║██╔══██╗
███████╗███████║███████║██████╔╝██╔████╔██║███████║
╚════██║██╔══██║██╔══██║██╔══██╗██║╚██╔╝██║██╔══██║
███████║██║  ██║██║  ██║██║  ██║██║ ╚═╝ ██║██║  ██║
╚══════╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝     ╚═╝╚═╝  ╚═╝
~$ Waterloo CompE · Full Stack SWE, QA, Data Engineering, LLM pipelines, AI Training, 3D/interactive · Won Canada's largest AI Hackathon.
▊
01ABOUT
CompE student at the University of Waterloo. I build full-stack AI systems, LLM pipelines, and the occasional thing that probably should not work, like multi-agent validation frameworks, RAG inference engines, and a 30M-parameter GPT trained from scratch in JAX. Engineering is a contact sport. Show up SCARRED.
languages
Python · Java · TypeScript · JS · PHP · SQL · C++
ml / ai
JAX · Keras · Optax · Orbax · RAG · Multi-Agent
frameworks
FastAPI · React · Node.js · Laravel · Selenium
infra
Docker · Azure · PostgreSQL · Redis · Supabase
02RESUME · git log
2026 → nowUW OrbitalGround Station DeveloperWaterloo
  • ›RESTful API for satellite command management with full HTTP status compliance and CORS middleware.
  • ›6-state command lifecycle: PENDING → SCHEDULED → ONGOING → COMPLETED / CANCELLED / FAILED.
  • ›100% telemetry uptime across ground station sessions.
Jan – Apr 26Hostalky (Practera)ML Research Fellowremote
  • ›End-to-end clinical documentation system with a 4-stage prompt pipeline.
  • ›9-agent validation framework targeting 5 defect categories: hallucination, negation, attribution, medication, omission.
  • ›Reduced hallucinations by 69.6% and improved documentation quality by 21.6%.
2025 → nowSample JunctionSoftware Dev & QA InternCanada
  • ›Full-stack CRUD app in Laravel MVC with 7 REST endpoints, MySQL + Eloquent ORM.
  • ›Selenium WebDriver + JMeter test suites raised coverage by 70%, validated to defect-free release.
  • ›Built an AI lead generation pipeline using Hunter.io + GPT to automate client outreach.
2025 → 2030University of WaterlooB.A.Sc. Computer EngineeringWaterloo, ON
03CERTIFICATIONS[4]
2026
Deep Learning Specialization
deeplearning.ai
verify ↗
2026
Microsoft Azure AI Fundamentals
AI-900
verify ↗
2026
Microsoft Azure Fundamentals
AZ-900
verify ↗
2026
Advanced Computing with NVIDIA C++
NVIDIA
verify ↗
04SELECTED WORK
// click any card to reveal details
04AI SCHOLARSHIP MATCHER
ScholarMatch
AI-powered scholarship platform. Gemini parses resumes; matching engine scores across 6+ student attributes against 100+ scraped opportunities. Auth0 + Vultr Docker deploy.
React · FastAPI · Supabase · PostgreSQL · Docker
100+ scholarships · 6+ attrs● production
ScholarMatch./ScholarMatch/preview2026100+ scholarships · 6+ attrs
↗ live↗ source
07AI ANATOMY COMPANION
Soma
Interactive 3D muscle-anatomy app. Users describe symptoms to Oso, an AI bear powered by Gemini, which highlights the relevant muscles on a voxel human model in real time. Multi-stage LLM pipeline: conversational chat → severity triage → per-muscle narration story → ElevenLabs TTS voiceover. 25 named muscle groups mapped to 3D mesh identifiers.
React · Vite · Express · Google Gemini · ElevenLabs · Three.js
25 muscles · 4-stage LLM pipeline · TTS● production
Soma./Soma/preview202525 muscles · 4-stage LLM pipeline · TTS
↗ live↗ source↗ devpost
05OTHER WORK[8]
// full archive — click any card to reveal details
01LLM INFERENCE OPTIMIZER
InferOpt
Won Canada's largest AI hackathon (100+ teams). Cuts inference cost up to 90% via template caching + slot filling, eliminating redundant tokens across repeated prompt patterns.
React · FastAPI · OpenAI Agents · Redis · Docker
−90% tokens · 1st / 100+● production
InferOpt./InferOpt/preview2026−90% tokens · 1st / 100+
↗ live↗ source
0230M-PARAM GPT, FROM SCRATCH
TinyGPT
Built a GPT-style transformer from scratch with multi-head self-attention, feedforward, positional encoding. Trained on tinyStories (100k stories, 3 epochs) on a T4.
JAX · Keras · Optax · Orbax · Grain
30M params · T4 · 3 epochs● open source
TinyGPT./TinyGPT/preview202630M params · T4 · 3 epochs
↗ live↗ source
034-STAGE PIPELINE + 9-AGENT QA
Hostalky
End-to-end clinical documentation system. 4-stage prompt pipeline + 9-agent validation framework targeting 5 defect categories (hallucination, negation, attribution, medication, omission).
Python · LLMs · React · FastAPI
−69.6% hallucinations · +21.6% quality● shipped
Hostalky./Hostalky/preview2026−69.6% hallucinations · +21.6% quality
↗ live⊘ proprietary
04AI SCHOLARSHIP MATCHER
ScholarMatch
AI-powered scholarship platform. Gemini parses resumes; matching engine scores across 6+ student attributes against 100+ scraped opportunities. Auth0 + Vultr Docker deploy.
React · FastAPI · Supabase · PostgreSQL · Docker
100+ scholarships · 6+ attrs● production
ScholarMatch./ScholarMatch/preview2026100+ scholarships · 6+ attrs
↗ live↗ source
05SATELLITE COMMAND API
UW Orbital
RESTful API for satellite command management with full HTTP status compliance, CORS middleware, and a 6-state lifecycle (PENDING → SCHEDULED → ONGOING → COMPLETED/CANCELLED/FAILED).
Python · FastAPI · SQLModel · React
6-state FSM · 100% telemetry● in flight
UW Orbital./UW Orbital/preview20266-state FSM · 100% telemetry
↗ live↗ source
06FINANCIAL ML & DATA ENGINEERING PIPELINE
PreCursor
End-to-end financial ML pipeline on Databricks. Medallion architecture (Bronze → Silver → Gold) ingesting 225K+ SEC Form 4 filings, FRED macro indicators, and 5 years of S&P 500 OHLCV data. Trains XGBoost and Temporal Fusion Transformer models on 42 engineered features. Streamlit dashboard surfaces insider trade signals, sector predictability, and daily model picks.
Python · Databricks · PySpark · XGBoost · PyTorch · Streamlit
613 stocks · 225K+ filings · 2 models● production
PreCursor./PreCursor/preview2025613 stocks · 225K+ filings · 2 models
↗ live↗ source
07AI ANATOMY COMPANION
Soma
Interactive 3D muscle-anatomy app. Users describe symptoms to Oso, an AI bear powered by Gemini, which highlights the relevant muscles on a voxel human model in real time. Multi-stage LLM pipeline: conversational chat → severity triage → per-muscle narration story → ElevenLabs TTS voiceover. 25 named muscle groups mapped to 3D mesh identifiers.
React · Vite · Express · Google Gemini · ElevenLabs · Three.js
25 muscles · 4-stage LLM pipeline · TTS● production
Soma./Soma/preview202525 muscles · 4-stage LLM pipeline · TTS
↗ live↗ source↗ devpost
08FULL-STACK + TEST AUTOMATION
Sample Junction
Full-stack CRUD app in Laravel MVC with 7 REST endpoints, MySQL + Eloquent. Selenium WebDriver + JMeter suites raised test coverage 70%, validated to defect-free release. Built an AI lead generation pipeline using Hunter.io + GPT to source clients from public data and automate outreach.
Laravel · MySQL · Selenium · Java
+70% coverage · 7 endpoints● shipped
./Sample Junction/preview2025+70% coverage · 7 endpoints
⊘ proprietary⊘ proprietary
06WRITING
2026
Training a 30M-param GPT from scratch
Multi-head self-attention, positional encoding, training on tinyStories with JAX / Optax / Orbax on a T4. What broke, what worked, what I'd do differently.
blog ↗
07CONTACT · ./hello
$ ./contact --reason="hire"
→ best inbox
paarths376@gmail.com
Reply window: 24h on weekdays. I read everything.
● grit-basedUTF-8LFv1.4.2 · 2026